Dario Amodei's essay "We Must Pace the Frontier" split the AI industry into three camps within days of publication, and that split is itself the most useful evidence about the state of AI governance. The Anthropic chief executive argues that AI capability is advancing faster than the ability to understand and control it, particularly as AI becomes capable of building the next generation of AI. His three-step plan — independent evaluators with employee-level access to frontier labs, shared safety standards among leading labs in democratic countries, and international agreements including with China — drew immediate support from Sam Altman, Elon Musk, Demis Hassabis and Ursula von der Leyen, and equally immediate rejection from Mark Zuckerberg and investor David Sacks. The speed of that disagreement matters more than any single endorsement.
What Amodei proposes and who signed on
Amodei's first step is the most concrete: give independent evaluators ongoing employee-level access to frontier labs to verify safety practices. Anthropic has said it is committing to this unilaterally, and OpenAI matched the embedded-evaluator pledge within hours of publication. The second step asks leading labs in democratic countries to align on shared safety standards, with government support where necessary. The third pursues international agreements, including with China, on the most dangerous capabilities. European Commission President Ursula von der Leyen backed the call to "pace the frontier" and said she would convene the major labs to discuss how Europe could support those efforts. Google DeepMind's Demis Hassabis said the essay points toward the right path forward.
The rejection camp made a narrower argument. Meta's Mark Zuckerberg rejected a coordinated slowdown, saying competition, legal liability and independent evaluation already give labs enough reason to build safely without anyone else setting their pace. Investor David Sacks put the same challenge more bluntly: if the companies calling for slower progress believe it is necessary, they can slow themselves down. A third group questioned the incentives behind the proposal. Investor Michael Burry called the slowdown talk self-serving, arguing it could benefit incumbent labs facing growing competition. None of the three camps disputed that AI capability is moving fast; they disputed who should decide what happens next.
The structural problem is that pacing asks companies competing for enormous economic rewards to collectively restrain themselves, while independent verification only asks them to expose more of what they are doing. Those are very different economic propositions. Verification can coexist with a market that rewards speed; collective restraint asks that market to work against its own incentives. Amodei's own essay, read closely, suggests a genuine global slowdown is the least likely of his proposed outcomes. Building is what scientists, engineers and entrepreneurs do, and if something looks possible, someone will find out whether it is.
What this means for companies deploying AI agents
For businesses adopting AI agents, the practical consequence is that the governance burden lands on the buyer, not on the lab. Deloitte's 2026 survey of more than 3,200 business and information technology leaders across 24 countries found that only 21% of organizations have mature governance for agentic AI. IBM found that 70% of technology executives say teams across their businesses are already deploying technology faster than IT can track, while only 11% said they were fully prepared for AI-agent deployment at scale. A small company can often absorb this by limiting agents to narrow, reversible tasks with a single owner. A large organization with many teams, tools and permissions cannot — its exposure grows with every integration it does not inventory.
What the news does not mean is that a coordinated slowdown is coming, or that any lab's voluntary pledge substitutes for oversight. METR's 2026 Frontier Risk Report shows Anthropic, Google, Meta and OpenAI gave METR access to their most capable internal models, including raw chains of thought, plus nonpublic information about internal use and monitoring. That is real progress, but access alone does not establish independence: who chooses the evaluator, who sets the standard, what the evaluator may disclose, and what happens when evaluator and company disagree all remain open. When selecting an AI vendor, the questions worth asking are who verified the system, against what standard, and what the vendor is obliged to report when something goes wrong.
The marker to watch is whether independent evaluation becomes a condition of doing business rather than a voluntary gesture. Aviation offers the cautionary case: two Boeing 737 MAX crashes in 2018 and 2019 killed 346 people, and the U. S. Department of Transportation Inspector General found Boeing and the Federal Aviation Administration had followed the established certification process. The process itself was not enough. If enterprise buyers begin requiring third-party verification as a procurement term, the governance gap starts to close from the demand side; if verification stays voluntary, deployment will keep outrunning the ability to govern it.
